OpenCode Go vs Zen: pricing, usage limits, and which models score best

OpenCode sells model access in two ways. OpenCode Go is a monthly subscription, from $10, for a set list of mostly open-weight models. Zen is pay-as-you-go: you add a balance and pay each model provider's token price, with Claude, GPT and Gemini included. Neither is needed to use OpenCode, which works with your own API keys too.

This guide compares the two using OpenCode's own docs and pricing, and Terminal-Bench scores published by Artificial Analysis.

OpenCode Go vs Zen: billing, models, limits and data

Both sell the cheap open models. On Go, you pay a fixed amount each month, and requests to a model stop when its allowance runs out. On Zen, each request is paid from your balance, and you can pick any model in the catalog, including Claude, GPT and Gemini.

OpenCode's documentation now calls the pay-as-you-go side "OpenCode Console", while its pricing page still calls it Zen. They are the same product.

Terminal-Bench 4.0 scores of models on OpenCode Go and Zen

Terminal-Bench gives a model real tasks in a terminal, such as fixing a build or setting up a server, and checks whether it finished them. The chart below shows the scores Artificial Analysis published for the models on each plan. Artificial Analysis ran them with its own agent, not OpenCode, so treat them as a measure of the model rather than of OpenCode.

The highest-scoring model in the chart is on Go, not Zen. Kimi K3.1 Pro solved 67.2% of tasks, ahead of Claude Sonnet 5.5 at 63.6% and Claude Opus 5.5 at 59.6%. Zen doesn't sell Kimi K3.1 Pro.

The catch is the allowance. OpenCode estimates that the $10 plan covers about 735 Kimi K3.1 Pro requests a month, and a single coding task can take dozens of requests. Most of the models on Go with many requests per month, like DeepSeek V4.1 Flash and Claude Haiku 5.5, score half as well or less.

OpenCode Go usage limits and requests per month, by plan

Go doesn't count requests. Each model has a monthly allowance in dollars, and every request is charged against it at that model's token price.

You can use 20% of a model's monthly allowance in any 5-hour window and 50% in any week, so a heavy day can't use up the month.

Monthly allowance and OpenCode's estimated requests per month, from OpenCode's Go docs on November 2, 2026.

• The allowances add up to far more than $10, but each applies to one model. You only get the full value if you spread work across models.
• Go Plus costs four times as much as Go but raises the limits unevenly: eight times on GLM-5.3, four times on most models, and twice on DeepSeek V4.1 Flash.
• OpenCode's estimates assume a typical request size for each model, such as about 77,850 tokens for Kimi K3.1 Pro, mostly from the prompt cache. Requests on a large codebase can be bigger, which means fewer requests per month.

OpenCode Go vs Zen token prices for the same model

Most models sold on both plans cost the same per token on both, including GLM-5.3, Qwen3.8 Max, Kimi K3, Claude Haiku 5.5, Grok 4.7, GPT 6 Luna and Gemini 4 Argon. The DeepSeek models are the exception.

Price per million tokens, from OpenCode's Go and Zen docs on November 2, 2026. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.

Off-peak, DeepSeek V4.1 Flash costs half as much on Go as on Zen. At peak hours, the prices are the same. DeepSeek V4 Pro is cheaper on Go off-peak, but at peak its output tokens cost more on Go than on Zen.

When to choose OpenCode Go, Zen or both

• Pick Go if you're happy working with open models and want a fixed bill. Use Kimi K3.1 Pro or GLM-5.3 for hard tasks, and DeepSeek V4.1 Flash, which OpenCode estimates covers about 300 times as many requests, for routine ones.
• Pick Zen if you want Claude Opus 5.5, Sonnet 5.5 or GPT 6.1 Sol in OpenCode, or you use OpenCode only now and then.
• Use both: OpenCode lets you connect several providers and switch per session, so you can use Go for routine work and a Zen balance for tasks that need a Claude or GPT model.
• Try the free models first. OpenCode lists about ten free models, which work only inside OpenCode. They're free for a limited time, and OpenCode's docs say some of them may use your prompts to improve the model, so don't use them on private code.

Running OpenCode on Go in the cloud

A Go key works wherever OpenCode does, including a cloud machine, so an agent can keep working when your laptop is closed. In Replicas, you connect an OpenCode Go key once and every OpenCode and Pi workspace uses it. You can add up to 10 keys from different Go subscriptions: when one reaches its limit, the next turn moves to the next key. Setup is in the OpenCode docs.

For other ways to run OpenCode, see OpenCode alternatives and OpenCode vs Claude Code.